On May 12, 2026, ElevenLabs shipped SDK v2.47.0. By the next day, LinkedIn was wall-to-wall posts shouting the same headline: claude-opus-4-7 just landed in the LLM enum. Three days in, a dozen agencies announced they were "already migrating their agents." And buried in the same changelog, in a paragraph nobody read to the end, another parameter slipped past everyone: keyterms.
Let me be direct. For the Quebec SMB running an AI voice agent in May 2026, switching LLMs will not fix your actual problem. Your actual problem is that your agent misunderstands when a customer says "I'm calling about Plomberie Latendresse" or "I need to speak with Madame Côté-Brassard." And no new reasoning brain is going to fix that.
Why the claude-opus-4-7 hype is misleading you
First, let's be honest about what ElevenLabs's LLM enum now contains: claude-opus-4-7, gpt-5.4, gpt-5.5, gpt-5.4-2026-03-05, gpt-5.5-2026-04-23, and qwen36-35b-a3b. Impressive on paper. And yes, for a complex query ("can I cancel my dental insurance and roll my deductible into next year?"), a stronger brain reasons better.
But here's what actually happens when you read SMB call transcripts in Quebec — we've reviewed more than 8,000 this year. The LLM almost never trips on reasoning. It trips on input. Scribe (ElevenLabs's speech-to-text module) hands the LLM "Place Saint-Brûle" instead of "Plomberie Latendresse." At that point, it doesn't matter whether the brain is claude-opus-4-7 or gpt-3.5 — it's answering a question nobody asked.
We already documented this brain-vs-input gap in our GPT-4o vs Claude 4.6 vs Gemini 3.1 Pro comparison from a few days ago. The takeaway fit in one line: LLM choice matters, but it matters far less than what you feed the LLM.
What keyterms actually does
Here's the verbatim line from the May 12, 2026 changelog: "The Scribe realtime WebSocket now accepts a keyterms parameter (array of strings, max 50 entries of up to 20 characters each) to bias the model toward specific terms."
Plain-English translation: you hand the agent a list of words it must recognize when spoken. Not force them in — if they aren't said, the agent doesn't invent them — just tilt the scale. The official keyterm prompting docs spell it out: it's contextual, not mechanical.
For the Quebec SMB, that breaks down into three wins, in order of impact.
1. Your business name stops getting mangled
"Boucherie Côté." "Garage Tremblay-Saint-Onge." "Esthétique Saint-Bruno." "Notaire Beauchemin & Associés." None of these names exist in Scribe's standard training distribution. The agent does its best — it guesses. With keyterms loaded, your name becomes a token the model actively reaches for when a customer says it, whether the accent is Saguenay-thick or Montreal-rapid.
2. Your products and services come out intact
If you're a vet clinic in Sherbrooke and your client calls about a "combined DAPP vaccine," a "Giardia fecal panel," or a "feline hepatic profile," those terms don't exist in casual speech. Before keyterms, your agent transcribed them however it felt. Now you put the 30-40 most-asked-for services in a list and the issue is gone.
3. Bilingual proper nouns stop glitching
This is the most-misunderstood trap in Quebec. A francophone client who says "j'ai un rendez-vous chez Bureau en Gros" and an anglophone who says "I have an appointment at Bureau en Gros" pronounce the name slightly differently. Without keyterms, the model leans toward the English phonetic interpretation in EN context, French in FR context — and regularly mangles bilingual brands like Couche-Tard, Jean Coutu, Familiprix, La Cage. A shared keyterms list across both languages stabilizes these names.
The counter-example that pushed me to write this
Last week we audited the AI voice agent of a notary firm in Lanaudière. The owner had just migrated to claude-opus-4-7 ("because it's better, right?"). His pain: 18% of calls ended with an annoyed client because the agent couldn't catch the name by ear.
We looked. The brain was flawless — it reasoned, it suggested, it routed correctly. But the input was broken. "Beauregard" became "beau regard." "Lavoie-Dansereau" became "la voix d'un sereau." The agent kept asking to spell, which made clients impatient.
We did two things. First, we put the LLM back on gpt-5.4 (cheaper, plenty of reasoning for this firm). Second, we loaded a keyterms list of 47 entries: the 12 notaries in the firm, the 18 municipalities they serve, and the 17 most-frequent legal terms. Additional monthly cost: $0. Setup time: 25 minutes.
72-hour result: the rate of calls where the agent asks the client to spell dropped from 18% to 3%. The owner emailed us the day before this article shipped: "Honestly, I assumed it was the LLM. It was a 600-byte text file."
What goes in your keyterms list — by sector
Here are the categories your keyterms file should cover, whatever your sector. The realtime limit is 50 entries of 20 characters each, so be surgical.
- Dental clinics: dentist and hygienist names, technical services ("endodontics," "Cerec sealant," "root planing"), regional insurers ("Croix Bleue Québec," "Industrial Alliance"). We covered the staffing-driven economics in detail in our piece on Quebec's 1,400 missing hygienists.
- Restaurants and hospitality: restaurant name, signature dishes with unique names, common allergies ("gluten," "lactose," "nut anaphylaxis"), events ("5 à 7," "Sunday brunch," "private room").
- Notary and law firms: partner names, mandate types ("protection mandate," "family patrimony," "incorporation"), local jurisdictions.
- Veterinary clinics: vet names, sensitive breeds ("French bulldog," "Australian shepherd"), common drugs ("Apoquel," "Bravecto," "Galliprant").
- Esthetics and hair salons: brand-specific treatments ("Olaplex," "HydraFacial," "microneedling"), in-house packages, esthetician names.
Simple rule: load proper nouns first (humans, brands), then rare technical terms, and only after that frequent words with awkward homophones. Keeping each entry under 20 characters forces concision — that's deliberate. A long compound ("decongestant-poultice") splits better into two separate entries.
Why this changes the unit economics in Quebec
Let's stay concrete. Looking at recent data on SMBs running AI voice agents, the pattern is clear: 80% of calls that go to voicemail are lost without callback. And the number-one mid-call abandonment cause when an AI picks up isn't, in fact, that clients refuse to talk to an AI. It's that they hang up after the third request to spell their name or repeat the service.
Every forced spelling is a leak. An internal study we ran at TECHMA in April 2026 across 23 Quebec SMBs showed that calls where the agent asks to spell ≥ 2 times have a 41% abandonment rate, versus 6% for calls with no spelling needed. Seven times more drop-off.
Now do the math. A typical Quebec SMB receives 80 calls per day. Say 25% contain a difficult proper noun (a family name, a compound business name, a product). Without keyterms, the agent stumbles on 60-70% of those entries. With keyterms tuned properly, that rate falls below 10% based on our tests on active accounts. That's roughly a dozen additional completed calls per day. Multiply by your average ticket.
What this won't fix (we're not naive)
Keyterms isn't magic. Three blind spots to keep in mind.
First, the parameter biases recognition but doesn't repair structural failures. If your phone microphone is bad or the client is on Highway 40 with the windows down, no keyterm will save the transcript. The real problem is elsewhere.
Second, the realtime cap of 50 entries × 20 characters is tight. A 12-notary firm with 18 municipalities and 17 legal terms already maxes out. Larger orgs will have to arbitrate or shift to batch mode (1,000 entries, but incompatible with a live agent).
Third — and this is the trap we see most often — many SMBs copy their keyterms list from a CRM export. That's wrong. The CRM holds the real written names; but in speech, customers say things. "9417654 Quebec inc." is in your CRM, but nobody phones in that way. Put the spoken usual name, not the legal one.
The ElevenLabs angle nobody's discussing
While we're on SDK v2.47.0, let's slip in another change nobody's mentioning: the same release added IP allowlisting on service account API keys. For a Quebec SMB under Law 25 and professional confidentiality, this is a real shift: you can now restrict a key's use to a single IP (the TECHMA Make.com server, say), closing an actual attack vector. Combined with ElevenLabs versioning — covered in our zero-downtime piece — you now have an agent that won't break on a Friday at 7 PM and won't fold to a stolen key.
Where keyterms protects quality, IP allowlisting protects security. Two legs of the same compliance story.
The honest case for claude-opus-4-7
Let's be clear before we wrap: claude-opus-4-7 isn't useless. It's an excellent brain for agents doing multi-step reasoning — managing a complex legal mandate, triaging a medical call with a deep decision tree, B2B price negotiation. For those cases, yes, pay the premium.
But for 80% of Quebec SMBs — the ones using a voice agent to book appointments, qualify prospects, transfer to the right human — gpt-5.4 or even claude-sonnet handles it. The difference shows up elsewhere: in transcription quality, in prompt consistency, in clean CRM integration. Not in the LLM.
That's why, when we configure a new agent at AgentiaVocal — and yes, we handle the full integration ourselves, from keyterms to Make.com — the keyterms list is the first thing we build with the client. Before we pick a voice, before we lock the prompt. Because an agent that doesn't hear properly is an agent that does nothing useful, regardless of how much you spent on its brain.
If you do only one thing this week
Open the transcript of your last 50 calls. Mark every instance where the agent asked the customer to repeat or spell. Note the word that caused it. In 80% of cases it will be: a family name, your business name, a product or service you offer, or a municipality you serve.
That's your initial keyterms list. No more than 50 entries, each under 20 characters. Load it into your ElevenLabs Scribe config. Measure again 72 hours later.
In one hour of work, you'll have done more for your AI voice agent's performance than every LLM migration of the past 12 months combined. And while your competitors keep debating claude-opus-4-7 on LinkedIn, your agent will simply understand the people calling you.
That, in May 2026, is the real competitive gap in Quebec.